← Glossary

Reading a study

The vocabulary for reading a study, placing the strength of a result and understanding what it adds in practice.

Randomized controlled trialRCT · Clinical trial#

Study in which participants are allocated at random between an intervention and a comparator. Random allocation is what allows a difference to be attributed to the intervention rather than to the characteristics of the groups. It is the only study type that permits speaking of an effect, and not of an association.

Double-blind trialBlinded trial · Masked trial#

Trial in which neither the participant nor the assessor knows who receives the treatment. The aim is to stop expectations from changing measurements and reported symptoms. A non-blinded trial on a subjective endpoint should be read with caution.

Crossover trialCrossover#

Trial in which each participant receives both interventions in turn, with an intervening period. It reduces variability between individuals and needs fewer participants. It does not suit treatments that leave a lasting effect.

Intention-to-treat analysisITT#

Analysis that keeps every participant in their group, including those who stopped the treatment. It reflects what the treatment produces in real life. An analysis that excludes dropouts can flatter the result.

Prospective cohort#

Follow-up of a group of people over time, with no assigned intervention. It describes associations and helps spot leads, not demonstrate causality: people who adopt one habit often adopt others, and statistical adjustment never corrects everything.

Case-control study#

Study comparing people with a disease to people without it, tracing back their past exposures. It is fast and useful for rare diseases. It is exposed to recall bias, since exposure is reconstructed after the fact.

Cross-sectional study#

Study measuring exposure and outcome at the same moment. It describes a snapshot of the population and cannot tell what precedes what. It is the study type behind most prevalence figures.

Ecological study#

Study comparing populations rather than individuals, for example a country's wine consumption and its mortality. It can suggest a lead but is exposed to the ecological fallacy: an association at group level does not hold for an individual.

Animal study#

Study conducted on an animal model, often the mouse. It is indispensable for understanding a mechanism, but its results do not transfer directly to humans, because of differences in metabolism, dose and lifespan. An effect on mouse longevity is not an effect in humans.

In vitro study#

Experiment performed on cells or tissues in the laboratory, outside the body. It serves to explore a mechanism and says nothing about the dose reachable in a person. Many claims about polyphenols rest on in vitro concentrations impossible to obtain through diet.

Meta-analysis#

Statistical combination of the results of several studies on the same question. Its value depends entirely on that of the included studies and on their heterogeneity: two serious meta-analyses can contradict each other when they do not include the same work or the same populations.

Absolute risk#

Number of additional or avoided events in a group, related to the number of people and to duration. It differs from relative risk, which can look spectacular for a rare event. Going from 2 to 1 case per 1,000 people is a halving in relative terms and 0.1 point in absolute terms.

Risk factor#

Characteristic associated with a higher probability of disease, without necessarily being a cause. Age, blood pressure and tobacco are examples of very different kinds. A risk factor can be useful for prediction without being useful for action.

Overall survivalOS#

Time elapsed until death, from all causes, in a trial. It is the most solid endpoint in oncology because it depends on no interpretation. Progression-free survival, which measures time before worsening, can improve without overall survival moving.

Systematic review#

Synthesis following an explicit method to find all studies on a question, assess them and summarize them, with or without meta-analysis. It is the method that best protects against selective reporting. A so-called narrative review, by contrast, chooses its references without a protocol.

Network meta-analysis#

Meta-analysis that indirectly compares several interventions with one another, even when no trial has opposed them directly. This is what makes it possible to rank training doses against each other, as in the 151 trials and 6,306 participants on exercise after 60.

Standardized mean differenceSMD · Effect size#

Effect size expressed in standard deviations, which allows comparison of studies that do not use the same measurement scale. It says how far one group differs from the other, without saying anything about the usefulness of the change for the person.

Hazard ratioHR#

Ratio between the risk of an event occurring in two groups over time. A ratio of 0.81 corresponds to a 19% lower risk in the treated group. It expresses a relative risk, and the baseline risk is needed to know what that changes in number of people.

Relative riskRR#

Ratio between the risk observed in an exposed group and that of an unexposed group. It is often presented alone, which exaggerates the importance of an effect when the baseline risk is low. A risk doubled for a rare event remains a rare event.

Odds ratioOR#

Ratio of odds, used in case-control studies and adjusted analyses. It approaches relative risk when the event is rare, and departs from it strongly when the event is frequent. Confusing the two is one of the most common reading errors.

Confidence interval95% CI#

Range of values in which the true estimate has a 95% chance of lying, under the assumptions of the model. The wider the interval, the less precise the result. An interval that crosses the neutral value means the effect cannot be distinguished from no effect.

P valuep-value · Significance threshold#

Probability of observing a result at least as extreme if the null hypothesis, that of no effect, is true. It does not give the probability that the hypothesis is true, nor the importance of the result. A p below 0.05 is a convention, not a proof.

Statistical power#

Probability that a study detects a real effect of the expected size. It rises with the number of participants and the size of the effect. An underpowered study can conclude that there is no effect when the effect exists.

Effect size#

Magnitude of a difference between two groups, expressed in the unit of the measure or in standardized units. It is distinct from statistical significance. A tiny effect can be significant in a very large sample.

Clinical significance#

Real importance of an effect for the person, distinct from its statistical significance. It is judged on the outcome that matters, independence or survival, not on an intermediate marker. It is the criterion most often missing from supplement claims.

Number needed to treatNNT#

Number of people to treat to avoid one additional event. It is calculated from absolute risk, not relative risk. An NNT of 50 means that 49 people receive a treatment without drawing individual benefit from it.

Selection bias#

Distortion that occurs when the people included in a study do not resemble those to whom the conclusion will be applied. It is common in volunteer cohorts, which are better educated and healthier than the general population. It limits the transfer of results.

Publication bias#

Tendency to publish positive results more often than negative ones. It artificially inflates the apparent effect of an intervention. Meta-analyses look for it with funnel plots and by comparison with trial registries.

Confounding factor#

Variable associated with both the exposure and the outcome, which can create a misleading association. For example, people who take vitamins see a doctor more often. Statistical adjustment reduces this bias but does not remove it when the variable is poorly measured.

Statistical adjustment#

Accounting in the model for variables likely to distort the observed association, such as age, tobacco or education level. It improves the reading but depends on the quality of the measurements. Adjustment does not turn a cohort into a trial.

Reverse causality#

Situation in which the observed effect actually precedes its supposed cause. Documented example: the association between physical activity and dementia disappears when activity is measured at least ten years before diagnosis, which suggests the disease was reducing activity rather than the opposite.

Placebo effect#

Improvement observed in people receiving a treatment with no active ingredient, linked to expectations and to the care context. It is real and often substantial on subjective symptoms, weaker on objective endpoints. This is why a trial without a control group proves nothing.

Nocebo effect#

Appearance of adverse effects in people receiving a product with no active ingredient, linked to anticipation. It explains part of the symptoms attributed to supplements and treatments. Recognizing it avoids blaming everything on the molecule.

Primary endpointPrimary outcome#

Measure chosen in advance to judge the effect of an intervention. It must be single or clearly ranked, and clinically relevant. Multiplying endpoints after the fact allows a significant result to be found by chance.

Surrogate markerIntermediate endpoint · Surrogate#

Biological measure used in place of the outcome that really matters, because it arrives sooner. An effect on a marker does not guarantee an effect on health: several supplements improve a marker without any trial showing clinical benefit.

Level of evidence#

Scale the School uses to place each recommendation: Established, Probable, Emerging or Experimental. It depends on the study type, the population and the outcome measured. A mechanism observed in the laboratory remains Experimental, even when it is appealing.

GRADE#

Method for grading the quality of evidence and the strength of recommendations, used by many learned societies. It accounts for risk of bias, inconsistency, imprecision and indirectness of the data. It is one of the systems that replaced simple levels A, B, C.

Preprint#

Manuscript circulated before peer review, often on a public server. It allows a result to be shared quickly, with no guarantee of quality. A preprint should not be used to change a recommendation before its reviewed publication and replication.

Peer review#

Assessment of a manuscript by researchers in the same field before publication. It filters obvious errors and methodological weaknesses, without guaranteeing the validity of the result. It is necessary but not sufficient.

Replication#

Reproduction of a study by an independent team, with comparable methods. It is the best barrier against a chance result or a laboratory effect. A finding that does not replicate must not ground a recommendation.

Conflict of interest#

Situation in which a financial or professional interest may influence the design, analysis or presentation of a study. It does not disqualify a piece of work, but it must be declared and read. Trials funded by the maker of the product tested are more often positive.

Dose-response relationship#

Variation of the effect with dose or duration of exposure. A regular and consistent relationship strengthens the causal hypothesis in observational studies. Its absence does not rule out an effect, but it weakens the interpretation.

Subgroup analysis#

Analysis restricted to part of the participants, by age, sex or risk level. It is useful when planned in advance and supported by a statistical interaction. Carved out after the fact, it easily produces misleading results.

Prevalence#

Proportion of people with a disease at a given moment. It depends on incidence and on the duration of the disease. A high prevalence can reflect longer survival rather than more new cases.

Incidence#

Number of new cases of a disease in a population over a given period. It measures the risk of onset, unlike prevalence, which photographs what exists. It is the indicator used in cohorts.